Method for annotation based on deep learning

4. The method of claim 3, wherein the training of the agent model based on the first bounding box to which the geometric transform is applied comprises training the agent model by using a reverse order of the shuffled actions as the ground truth (GT) of the action learned by the agent model based on the first bounding box to which the geometric transform is applied.

8. The method of claim 7, wherein the completion condition of the learning based on the first constraint comprises a case of maintaining an accumulation value of the compensation calculated through the learning based on the first constraint to be greater than or equal to a predetermined reference value.

10. The method of claim 2, wherein the actions selectable by the agent model comprise movement of the bounding box, width adjustment of the bonding box, height adjustment of the bounding box, and angle transform of the bounding box.

11. The method of claim 1, wherein the constraint is a condition limiting the geometric transform of the second bounding box to be performed within a predetermined range determined based on the first bounding box.

13. The method of claim 5, wherein the actions selectable by the agent model comprise movement of the bounding box, width adjustment of the bonding box, height adjustment of the bounding box, and angle transform of the bounding box.

14. The method of claim 7, wherein the actions selectable by the agent model comprise movement of the bounding box, width adjustment of the bonding box, height adjustment of the bounding box, and angle transform of the bounding box.

15. The method of claim 1, wherein the step of training the agent model based on the second bounding box randomly sampled based on the second constraint is a retraining of the agent model performed after completion of the step of training the agent model based on the second bounding box randomly sampled based on the first constraint.

18. The computer program of claim 16, wherein the operation of training the agent model based on the second bounding box randomly sampled based on the second constraint is a retraining of the agent model resulting from the operation of training the agent model based on the second bounding box randomly sampled based on the first constraint.

21. The computing device of claim 19, wherein the processor is configured by the computer program codes to perform the training of the agent model based on the second bounding box randomly sampled based on the second constraint as a retraining of the agent model after performing the training of the agent model based on the second bounding box randomly sampled based on the first constraint.